CN-122020162-A - Training method and detection method of aeroengine performance detection model

CN122020162ACN 122020162 ACN122020162 ACN 122020162ACN-122020162-A

Abstract

The application provides a training method and a detection method of an aeroengine performance detection model, wherein the training method comprises the steps of generating covariance matrixes for representing the cooperative variation degree among different sensor characteristics according to a plurality of sensor characteristics collected by a sensor of each initial training sample in an initial training set; based on covariance matrix, performing feature selection on a plurality of sensor features acquired by sensors of a plurality of initial training samples to obtain a target training set, inputting the target training samples to an initial prediction model for each target training sample, and outputting performance prediction information of the aeroengine, wherein the initial prediction model is constructed according to a transform network and an attention network, and based on the performance prediction information and performance labels of the target training samples, adjusting model parameters of the initial prediction model to obtain a trained aeroengine performance detection model.

Inventors

HU CHUNYAN
SHEN YAFENG
ZHOU MINGYANG
ZENG QINGWEN
HAN BO
KANG FANG

Assignees

中国科学院工程热物理研究所

Dates

Publication Date: 20260512
Application Date: 20260120

Claims (10)

1. A method of training an aircraft engine performance test model, comprising: Generating a covariance matrix representing the degree of cooperative variation among different sensor characteristics according to a plurality of sensor characteristics acquired by a sensor of each initial training sample in an initial training set, wherein the different sensor characteristics represent different performance detection parameters of corresponding different components in the aeroengine; Based on the covariance matrix, performing feature selection on a plurality of sensor features acquired by sensors of a plurality of initial training samples to obtain a target training set, wherein the target training samples in the target training set comprise a plurality of sensor features determined through feature selection, and the determined sensor features represent key features of engine thermodynamics, mechanical dynamics and aerodynamics; Inputting the target training samples into an initial prediction model for each target training sample, and outputting performance prediction information of the aeroengine, wherein the initial prediction model is constructed according to a transform network and an attention network; And adjusting model parameters of the initial prediction model based on the performance prediction information and the performance label of the target training sample to obtain a trained aeroengine performance detection model.
2. The method of claim 1, wherein generating a covariance matrix that characterizes a degree of collaborative variation between different sensor features from a plurality of sensor features acquired by the sensors for each initial training sample in the initial training set comprises: generating an average training sample according to the initial training samples, wherein the average training sample comprises a plurality of average sensor characteristics, and the average sensor characteristics are determined according to sensor characteristics corresponding to the average sensor characteristics in the initial training samples; and generating the covariance matrix according to the number of the average training samples and the initial training samples.
3. The method of claim 1, further comprising, prior to generating the covariance matrix: and based on the Pearson correlation coefficient, carrying out feature screening on the sensor features in the initial training samples to obtain a plurality of processed initial training samples.
4. A method according to claim 3, wherein feature screening the sensor features in the plurality of initial training samples based on pearson correlation coefficients to obtain a plurality of processed initial training samples comprises: calculating a correlation coefficient between the target feature and the aeroengine performance according to the target features in the initial training set and the performance labels corresponding to each initial training sample aiming at any target feature in the sensor features in the initial training samples; And deleting the target characteristic of each initial training sample in the initial training set under the condition that the correlation coefficient is smaller than a correlation threshold value to obtain the processed initial training samples.
5. The method according to claim 1 or 2, wherein feature selection of a plurality of sensor features acquired by the sensor for a plurality of the initial training samples based on the covariance matrix results in a target training set, comprising: Calculating a projection variance according to the covariance matrix for any one of a plurality of sensor features; Calculating a unit feature vector according to the covariance matrix and the projection variance; And generating the target training set according to the plurality of unit feature vectors and the initial training sample.
6. The method of claim 5, wherein generating the target training set from a plurality of unit feature vectors and the initial training samples comprises: Generating an average training sample according to the initial training sample; Selecting a plurality of target feature vectors from a plurality of unit feature vectors according to the sizes of the unit feature vectors, and constructing a transformation matrix based on the plurality of target feature vectors; And generating the target training set according to the transformation matrix and the average training sample.
7. The method of claim 1, further comprising, prior to inputting the target training sample into an initial predictive model: Aiming at any sensor characteristic in the target training set, processing the sensor characteristic and random noise by utilizing a generated countermeasure network to obtain a discrimination parameter; Generating feature distribution information according to the target discriminator obtained by carrying out deviation solving on the discrimination parameters and the sensor features; generating a data enhanced target training set based on feature distribution information corresponding to different sensor features; Wherein prior to generating the data-enhanced target training set based on the feature distribution information, further comprising: And carrying out genetic optimization on the characteristic distribution information by utilizing a genetic algorithm to obtain optimized characteristic distribution information.
8. The method of claim 1, wherein inputting the target training samples into an initial predictive model and outputting performance prediction information for the aircraft engine comprises: Generating a dimensional space matrix according to the embedded weight matrix of the transform network and the target training sample; For any attention head in the attention network, respectively calculating a query vector, a key vector and a value vector according to the dimensional space matrix so as to calculate attention head weights according to the query vector, the key vector and the value vector; splicing the attention head weights to obtain fusion characteristics; And inputting the fusion characteristics into a feedforward network, and outputting the performance prediction information.
9. The method as recited in claim 1, further comprising: Generating an individual chromosome according to a plurality of model parameters of the aeroengine performance detection model, and generating an initial population based on a value range of each model parameter in the individual chromosome, wherein the initial population comprises a plurality of initial individuals; The following operations are performed iteratively: operation 1, inputting a plurality of test samples into an aeroengine performance detection model comprising the initial individuals aiming at each initial individual, and outputting performance test information corresponding to each test sample; operation 2, calculating individual fitness of the initial individual according to a plurality of performance test information and the test labels of each test sample; An operation 3, selecting a target individual from a plurality of initial individuals based on the individual fitness to perform genetic operation, and taking the individual obtained by the genetic operation as a new initial individual to return to the execution of the operation 1; And under the condition that the individual fitness is smaller than a preset fitness threshold value, carrying out parameter updating on the aeroengine performance detection model based on an initial individual corresponding to the individual fitness to obtain an updated aeroengine performance detection model.
10. A method for detecting the performance of an aircraft engine, comprising: Acquiring a plurality of performance parameters to be detected, which are acquired by a target sensor in an aero-engine; Inputting a plurality of the performance parameters to be detected into an aeroengine performance detection model, and outputting predicted performance information, wherein the aeroengine performance detection model is trained based on the method of any one of claims 1 to 9.

Description

Training method and detection method of aeroengine performance detection model Technical Field The application relates to the technical field of aeroengines, in particular to a training method and a detection method of an aeroengine performance detection model. Background Aeroengines are the core propulsion systems of aircraft, the operating conditions of which are directly related to flight safety and operating efficiency. Therefore, the method has important engineering significance for accurately monitoring the state of the engine and managing the health of the engine. Currently, engine health monitoring systems have been widely used in airlines and maintenance, repair and overhaul services. If the state of key parameters of the engine such as thrust, residual service life and the like can be estimated and predicted in real time, the system is beneficial to realizing an active maintenance strategy, such as the replacement of key components (such as turbine blades and sensors) or the calibration of the system is arranged according to the state prediction result, so that the unplanned off-time is effectively reduced, and the running safety is improved. In the related technology, when performance parameters such as thrust, residual service life and the like of an aero-engine are predicted, the problems of low prediction accuracy and more occupied computing resources exist. Disclosure of Invention In view of the above, the application provides a training method and a detection method for an aeroengine performance detection model. One aspect of the application provides a training method of an aeroengine performance detection model, comprising the following steps: Generating a covariance matrix representing the cooperative variation degree among different sensor characteristics according to a plurality of sensor characteristics collected by a sensor of each initial training sample in an initial training set, wherein the different sensor characteristics represent different performance detection parameters corresponding to different components in the aeroengine, selecting the characteristics of the sensors collected by the sensors of the plurality of initial training samples based on the covariance matrix to obtain a target training set, wherein the target training samples in the target training set comprise a plurality of sensor characteristics determined through characteristic selection, the determined sensor characteristics represent key characteristics of engine thermodynamic, mechanical dynamics and aerodynamics, inputting the target training samples into an initial prediction model aiming at each target training sample, outputting performance prediction information of the aeroengine, wherein the initial prediction model is constructed according to a transform network and an attention network, and adjusting model parameters of the initial prediction model based on the performance prediction information and performance labels of the target training samples to obtain a trained aeroengine performance detection model. The application further provides a method for detecting the performance of the aeroengine, which comprises the steps of acquiring a plurality of performance parameters to be detected, which are acquired by a target sensor in the aeroengine, inputting the performance parameters to be detected into a performance detection model of the aeroengine, and outputting predicted performance information. The application provides a training device of an aeroengine performance detection model, which comprises a generation module, a prediction module and a selection module, wherein the generation module is used for generating a covariance matrix representing the cooperative variation degree among different sensor characteristics according to a plurality of sensor characteristics acquired through sensors of each initial training sample in an initial training set, the different sensor characteristics represent different performance detection parameters corresponding to different components in the aeroengine, the selection module is used for carrying out characteristic selection on the sensor characteristics acquired through sensors of the plurality of initial training samples based on the covariance matrix to obtain a target training set, the target training samples in the target training set comprise a plurality of sensor characteristics determined through characteristic selection, the determined sensor characteristics represent key characteristics of engine thermodynamic, mechanical dynamics and aerodynamics, the prediction module is used for inputting the target training samples into an initial prediction model for each target training sample, outputting performance prediction information of the aeroengine, the initial prediction model is constructed according to a tranm network and an attention network, and the adjustment module is used for carrying out characteristic adjustment on the target training samples